Francesco Galletta and Felippe Felisola Caso demoed a financial-analyst harness designed to earn trust before it calculates anything, using a wedding budget as the live example in a workshop carried by AI Engineer. The system they showed runs local finance math through tools and scripts instead of letting a language model do arithmetic, with the explicit goal of handling fixed and variable costs, guest counts and staged payments without hand-waving.
Trust was the brief.
Galletta, who works on data at Monzo with Felisola Caso, opened on that bottleneck. He said speed is no longer the constraint, with background agents already running, and argued the gap is how a system connects with a user when money is at stake. He laid out five design principles that stack in order: shared domain to agree on terms, consistency via tool calling, evidence that shows receipts, accountability for what a swarm does, and verification from a separate point of view. He also said the team would not walk through a product readme or pitch a specific technology, and that the harness on display was intentionally simple.
Felisola Caso, who described his path from electrical engineering through data studios and dashboards to building internal AI-driven productivity tools for Monzo's data team, framed the same problem from an adoption angle. He said Monzo's adoption is deliberate, not token maxing, and that the value has to be concrete before spend scales. He said he and Galletta were not speaking for Monzo in the wedding example, but applying lessons from Monzo to a neutral use case, and he pointed to a separate breakout on Agent Chip, the bank's in-house remote agent capability. Where Galletta stressed structure, Felisola Caso stressed restraint, and neither claimed the model presented was a silver bullet. Monzo, which describes itself as building a smart bank on your smartphone, has grown from challenger to scaled operator, and the pair placed their approach against that backdrop rather than against feature velocity.
That positioning matters because the most common alternative in the wild is still a chat window over a sheet. The workshop's harness is the opposite bet: a constrained workspace with a repo, a small UI and a blank-canvas chat that forces the user to define the goal, then pushes back with questions. In the demo, the empty state invites a prompt like planning a budget or an investment, while the worked wedding example shows the fuller flow. The contrast is not with a single named product but with the default pattern of bolting an OpenAI client onto every data connection and hoping for the best, a pattern both speakers dismissed.
They were also candid about limits. Galletta warned something will break live and said parts are overengineered for illustration, not as a polished product or UX reference. Felisola Caso said building is not the hard part, knowing how to approach the problem is, and he invited pushback on whether trust is overrated. Neither offered metrics for when trust is sufficient, nor a concrete evaluation suite beyond traces, loops, unit tests and observability as components. The repeated caution was that verification cannot be an afterthought in finance, or results skew to the user's biases.
What they declined to define is the core tension they introduced. If accountability stays with the person who runs 100 agents, how is that accountability exercised without reviewing every diff, and what independent source counts as verification when the user's own numbers are the ground truth they distrust.
